The national Graduate School of Language Technology (GSLT) Course: Lexical semantics Åke Viberg 2002-09-12 Lecture 1. Lexical semantics: An overview (1) Dictionary vs. Lexicon (Handke 1995) Dictionaries: word-stores primarily consulted for information (printed or electronic) Lexicons: word-stores constituting a component within NLP systems (in brains or computers) (2) Two approaches to lexical semantic analysis 1. Semantic relations 2. Lexical decomposition (3) Major semantic lexicons on computer (1) Princeton WordNet (G A Miller; based on semantic relations) (2) FrameNet (based on Fillmore s Frame semantics) (3) SIMPLE (based on Pustejovsky s Generative lexicon) (4) Methodology for building computational lexicons (1) Restructuring of existing dictionaries LDOCE (Boguraev & Briscoe 1989) Bilingual dictionaries (Fontenelle 1997) (2) Corpus-based analysis (a) The WASP bench (Adam Kilgariff: http://www.itri.brighton.ac.uk/peopleindex.html) (b) Translation corpora (Stig Johansson) (5) Models of the Mental/Neural lexicon Levelt s psycholinguistic model Jackendoff s Conceptual semantics 1
WORLD N MIND Visual system O O B I Language Conceptual Spatial Haptic system J S Structure (CS) Structure E E (SpS) Action system C S meaning T S Conceptual Structure Spatial Structure Judgements and inferences having to do with Predicate-argument relations Category membership The type-token distinction Quantification Exact shapes Locations Forces The place of meaning according to Jackendoff (Foundations of Language. 2002) (6) Universality vs. Linguistic relativity (a) Universality: Wierzbicka s Natural Semantic Metalanguage (NSM) The basic purpose of NSM is to establish a minimal set of lexical concepts which can serve to paraphrase all the words in any human language. Table 1. The 37 primitives of the Natural Semantic Metalanguage (NSM) Substantives Mental predicates Determiners/Quantifiers Actions/Events Meta-predicates Time/Place Partonomy/Taxonomy Evaluators/Descriptors I, YOU, SOMEONE, SOMETHING, PEOPLE THINK, SAY, KNOW, FEEL, WANT THIS, THE SAME, OTHER, ONE, TWO, MANY, ALL DO, HAPPEN NO, IF, CAN, LIKE, BECAUSE, VERY WHEN, WHERE, AFTER, BEFORE, UNDER, ABOVE HAVE PARTS, KIND OF GOOD, BAD, BIG, SMALL 2
envy X feels envy. = X feels like someone who thinks this: something good has happened to that other person it hasn't happened to me I want things like that to happen to me lie X lied to Y. = X said something to Y X knew it was not true X said it because X wanted Y to think it was true [people would say: if someone does this it is bad] The system shown in Table 1 represents an earlier version. For a more recenr version, see Wierszbicka (1996). (b) Stephen Levinson: Linguistic relativity and the mapping problem The mapping problem: How can the child learn to map word forms unto word meanings? Levinson (2001) distinguishes three distinct degrees of complexity with respect to the mapping problem: Degree 1: mapping language-specific word forms unto common, universal semantic units Degree2: mapping word forms unto language-specific meanings constructed from universal concepts Degree 3: mapping word forms unto language-specific word meanings, given nonuniversal working concepts 3
(7) Accounting for the meaning(s) of a word (a) The problem of semantic representation 1) Necessary and sufficient conditions 2) Prototypes 3) Underspecification (b) Polysemy 1. Sense enumeration 2. Apresjan: Regular polysemy 3. Cognitive linguistics Prototypes + Regular semantic extensions Langacker s usage-based model 4. Pustejovsky s Generative lexicon Underspecified representations + Generative devices 4
(8) Pustejovsky s Generative Lexicon (a) Four levels of lexical representation (i) ARGUMENT STRUCTURE (ii) EVENT STRUCTURE (iii) QUALIA STRUCTURE (iv) LEXICAL INHERITANCE STRUCTURE Lexical representations may be underspecified. (b). A set of generative devices These devices connect the four levels and provide for the compositional interpretation of words in context (i) Type coercion Mary enjoyed the movie. Mary enjoyed watching the movie. (ii) Co-composition Mary baked a cake. Mary baked a potato. (iii) Selective binding a fast boat / a fast typist ; a sad man/ a sad day 5
(9) Qualia structure Pustejovsky/SIMPLE EuroWordNet FORMAL FORM Taxonomic relationships CONSTITUTIVE COMPOSITION a) Sensory attributes b) Material c) Part-Whole structure AGENTIVE ORIGIN Artifact: Manner of production Organism: Life-cycle stages TELIC FUNCTION Artifact: Typical use People a) Occupation: teacher b) Characteristic activity: miser c) Current activity: pedestrian (10) Dot objects Printed matter: [PHYSICAL OBJECT n INFORMATION] Physical object Eno the cat is sitting on yesterday s New York Times. Information Yesterday s New York Times really got me upset. Physical object & Information That book about the Crimean War has a red cover. Organization The newspaper has just fired its sports editor. 6
(11) Event structure: (i) The specific events and their type (ii) The ordering restriction over these events (a) Event types: state, process, transition (b) Head: e* : The head = the most prominent subevent in the event structure of a predicate (c) Temporal ordering relations (ordering restriction RESTR) < α :exhaustive ordered part of Ex: build o α : Complete simultaneity. Two completely simultaneous subevents Ex: accompany < o α : ordered overlap. an event containing two subevents, e 1 and e 2, where e 1 starts before e 2 Ex: walk (e 1 = motion of legs, e 2 = motion of body) 7
drive EVENTSTR = ARGSTR = QUALIA E 1 = e 1 :process E 2 = e 2 :process RESTR = < o ARG1 = x:human ARG2 = y:vehicle FORMAL = move(e 2,y) AGENTIVE = drive_act(e 1,x,y) arrive EVENTSTR = E 1 = e 1 :process E 2 = e 2 :state RESTR = < HEAD= e 2 ARGSTR = ARG1 = x:ind D-ARG1 = y:location QUALIA= FORMAL = at(e 2,x,y) AGENTIVE = arrive_act(e 1,x) 8